Accurate and timely intraoperative classification of vocal cord leukoplakia by machine learning assisted handheld OCT

Precise and timely intraoperative classification of vocal cord leukoplakia is vital for early staging and detection of laryngeal cancer. Current endoscopic imaging, such as narrow-band imaging (NBI) and white-light laryngoscopy (WLI), has the capability to evaluate the malignancy, but has low sensitivity (<70%) for grading the low-risk dysplasia. For the vocal fold lesions, optical coherence tomography (OCT), a non-invasive imaging technology, has been widely used to show the texture and pixel change of lesions caused by squamous epithelial cell hyperplasia and atypical hyperplasia. Here, a handheld OCT probe (OD = 3 mm, length = 25 cm) was designed and fabricated, and a swept-source OCT system was built for vocal lesion monitoring during the surgery. Home-built OCT probe has both a small OD and a long length for penetration depth imaging with good quality, which may enable the accurate and timely intraoperative localization detection of vocal cord tissues. The probe system has a 2D imaging speed of 100 frames per second, 7.96 µm lateral resolution, 16.24 µm axial resolution, and an imaging depth of 2.12 ± 0.43 mm in vocal fold tissues. We collected 378 OCT b-scan images (114 low risk, 175 high risk, 89 malignant) from 12 patients (4 low risk, five high risk, three malignant). A total of eight features of single B-scan images were extracted, including the mean pixel intensity (PI), median PI, minimum PI, and maximum PI; meanwhile, the mean attenuation coefficient (AC), median AC, minimum AC, and maximum AC of A-lines. The random forest (RF) model achieved the best overall accuracy (92.59%) and recall (93.25%) for low risk, high risk, and malignant classification, using 5-fold validation, with each patient's assessment completed within 10 minutes. Our OCT system was demonstrated to provide rapid and accurate pathological diagnosis of vocal cord leukoplakia intraoperatively. Our handheld OCT system facilitates identification of low-risk dysplasia with a sensitivity of up to 84.21% during the surgery, offering critical guidance on whether to opt for conservative or surgical treatment based on the lesion's severity.